By Guest contributors |
Early AI adoption tinkered at the edges, resulting in faster task completion within the same workflows. While this may have seemed like transformation, it wasn’t.
Now, AI agents are shifting the ground. These systems don’t just assist, they act.
The question is no longer ‘where can I apply AI?’. It’s ‘what does this function become when AI is the decision-maker, and not just the assistant?’
At ZS, we’re using Azure AI to test this new type of question in supply chain and manufacturing. The goal is to help pharmaceutical companies get medicines to patients more quickly and with higher quality.
We’ve developed agents for the two key ways companies monitor product safety and quality as part of the manufacturing process.
The first is in near real-time monitoring of manufacturing performance across critical key performance indicators such as yield, defects and deviations. This step helps teams to spot and resolve issues as they happen.
The second is in batch disposition activities to confirm that final manufacturing results and supporting documentation meet compliance standards and fall within allowable limits. These steps are intended to help ensure every dose is safe, effective and meets strict quality standards.
To further streamline these safety steps, we’ve developed a coordinated system of AI agents. A planning agent oversees the processes handled by four specialised agents, ensuring each one stays focused on its core task.
First, our data aggregator agent pulls in data from multiple manufacturing data systems and detects anomalies. It also handles data cleaning and exception management, laying the foundation for accurate analysis.
Then, once the data is clean, our batch conformance agent compares it against expected standards, uses batch profiling to flag inconsistencies and generates compliance summaries to guide downstream decisions.
Meanwhile our review management agent makes sure that the right people are looped in at the right time. It triggers review notifications, packages historical quality data and manages approval via application programming interfaces and email workflows.
Finally, the quality assurance (QA) assist agent recommends disposition paths – either release, hold or rework – based on the evidence. This is where AI becomes truly assistive and, while human QA still has the final say, the agent ensures decisions are rooted in data and are compliant with standard operating procedures.
Let’s return to the question of what does this function become when AI is the decision-maker, and not just the assistant.
With these agents, teams get AI-derived insights and actionable recommendations directly, instead of chasing down documents or manually cross-checking reports. This means that quality escapes and recalls become less frequent as AI agents detect subtle anomalies and ensure every batch meets policy standards.
Vikas Hegde is principal and Satish Jha is associate principal at ZS